setup: drop the PAC perf patch

Earlier PACT versions used a patched perf to feed sampling data over shared
memory. This version samples PEBS and CHA/uncore counters directly through
perf_event_open and no longer needs it, so remove setup/perf/ and rescope the
README, modeling scripts, LICENSE, NOTICE, and .gitignore references to it.

Author: huaicheng

.gitignore

@@ -9,9 +9,6 @@ src/pact
 *.txt
 run/results/
 
-# perf build tree produced by setup/perf/install-perf.sh
-setup/perf/linux-5.15-pact/
-
 # Editor / OS cruft
 .DS_Store
 *.swp

LICENSE

@@ -31,9 +31,9 @@ licenses; see the NOTICE file for the full list. In particular:
     original licenses. See NOTICE for provenance and license of each.
 
   * Linux kernel patches under baselines/ and setup/ (the kernel modules under
-    setup/kernel/ and the perf patch under setup/perf/) are derivative works of
-    the Linux kernel and are therefore governed by the GNU General Public
-    License, version 2 (GPL-2.0), the license of the kernel they patch.
+    setup/kernel/) are derivative works of the Linux kernel and are therefore
+    governed by the GNU General Public License, version 2 (GPL-2.0), the license
+    of the kernel they patch.
 
   * Documentation, figures, and data files are additionally available under
     the Creative Commons Attribution 4.0 International License (CC BY 4.0).

NOTICE

@@ -21,8 +21,8 @@ License:    Public Domain or MIT-0 (No Attribution), at your option
 ------------------------------------------------------------------------------
 
 Linux kernel patches and out-of-tree kernel modules under baselines/ and
-setup/ (the kernel modules under setup/kernel/ and the perf patch under
-setup/perf/) are derivative works of the Linux kernel and are licensed under
-GPL-2.0, matching the kernel they modify. The upstream tiering systems they
-reproduce (TPP, Memtis, Colloid, NBT, Nomad, Soar/Alto) are the work of their
-respective authors; see each subdirectory's README for provenance.
+setup/ (the kernel modules under setup/kernel/) are derivative works of the
+Linux kernel and are licensed under GPL-2.0, matching the kernel they modify.
+The upstream tiering systems they reproduce (TPP, Memtis, Colloid, NBT, Nomad,
+Soar/Alto) are the work of their respective authors; see each subdirectory's
+README for provenance.

README.md

@@ -63,8 +63,7 @@ PACT/
 ├── src/             # the PACT runtime (C source + Makefile)
 ├── setup/           # host bring-up - do this first
 │   ├── kernel/      #   build/boot Linux 6.3 + the tierinit & kswapdrst modules
-│   ├── env/         #   uncore-freq pinning, CXL/NUMA layout, governor, THP/KSM off
-│   └── perf/        #   build the PAC-patched perf used for PAC sampling
+│   └── env/         #   uncore-freq pinning, CXL/NUMA layout, governor, THP/KSM off
 ├── run/             # run PACT on a workload (run-pact.sh, workloads.sh)
 ├── baselines/       # SOTA systems we compare against (TPP, NBT, Nomad, Colloid, Memtis, Soar/Alto)
 └── modeling/        # PAC modeling scripts (Figure 3)
@@ -94,8 +93,9 @@ The artifact targets the same class of machine used in the paper:
   [`setup/kernel/`](setup/kernel/)). Baselines use their own kernel versions -
   see [`baselines/`](baselines/).
 - **Software:** `gcc`, `make`, `libnuma`, `numactl`, `vmtouch`, `gnuplot`
-  (plotting), and a patched `perf` built by
-  [`setup/perf/install-perf.sh`](setup/perf/install-perf.sh).
+  (plotting), and `perf` (the PAC modeling scripts in [`modeling/`](modeling/)
+  read counters via stock `perf stat`). The PACT runtime samples the PMU
+  directly through `perf_event_open` and needs no external `perf` binary.
 - **Privileges:** kernel install, module loading, and the environment-prep
   scripts require root.
 
@@ -215,7 +215,8 @@ constraints of the current implementation are worth knowing up front:
   CXL-like slow tier on a 2-NUMA-node host. More than two tiers, multi-socket
   fan-out, and multi-node setups are out of scope for this release.
 - **Hardware sampling required.** PAC profiling depends on PEBS plus CHA/uncore
-  occupancy counters and a PAC-patched `perf`, and the runtime must run as root.
+  occupancy counters (read directly via `perf_event_open`), and the runtime must
+  run as root.
 - **Attaches to a running, pre-pinned workload.** PACT manages an existing
   process by PID; it does not launch or CPU-pin workloads itself (the `run/`
   scripts handle that around it).

modeling/README.md

@@ -159,7 +159,7 @@ no need to add `numactl` inside the command.
 
 - `numactl` - `apt install numactl`
 - `vmtouch` - `apt install vmtouch` (GAPBS graph preloading)
-- `perf` - patched perf built by `setup/perf/install-perf.sh`; override its location with `PERF=/path/to/perf ./run-models.sh` (defaults to `perf` in `$PATH`)
+- `perf` - stock `perf` from `linux-tools`; the four counters are read from `perf stat` text output. Point at a specific build with `PERF=/path/to/perf ./run-models.sh` (defaults to `perf` in `$PATH`)
 - `gnuplot` - required only for plotting, not for data collection
 - Workload binaries installed and paths configured in `run-models.sh`. The
   GAPBS runs use the same `kron.sg` graph as `run/` - generate the paper's

modeling/run-models.sh

@@ -15,8 +15,8 @@ LOCAL_NODE=0    # Tier 1: local DRAM   (~90ns)
 NUMA_NODE=1     # Tier 2: remote NUMA  (~140ns)
 CXL_NODE=2      # Tier 3: CXL          (~190ns)
 
-# Patched perf binary built by setup/perf/install-perf.sh.
-# Override with: PERF=/path/to/perf ./run-models.sh   (default: perf in PATH)
+# Stock perf is enough — the four counters below are read from perf stat text
+# output. Override with: PERF=/path/to/perf ./run-models.sh (default: perf in PATH)
 PERF="${PERF:-perf}"
 
 # Four perf counters; all collected in a single perf stat pass